high performance computing on graphics processing units: hgpu.org

We introduced four prototypes of General Purpose GPU solutions by Compute Unified Device Architecture (CUDA) on NVidia GeForce 8800GT and Tesla C870 for a practical Curved Ray Prestack Kirchhoff Time Migration program, which is one of the most widely adopted imaging methods in the seismic data processing industry. We presented how to re-design and re-implement the original CPU code to efficient GPU code step by step. We demonstrated optimization methods, such as how to reduce the overhead of memory transportation on PCI-E bus, how to significantly increase the kernel thread numbers on GPU cores, how to buffer the inputs and outputs of CUDA kernel modules, and how to utilize the memory streams to overlap GPU kernel execution time, etc., to improve the runtime performance on GPUs. We analyzed the floating point errors between CPUs and GPUs. We presented the images generated by CPU and GPU programs for the same real-world seismic data inputs. Our final approach of Prototype-IV on NVidia GeForce 8800GT is 16.3 times faster than its CPU version on Intel’s P4 3.0G.